Channel Estimation Method, Device, Equipment and Medium for Large-Scale MIMO System Based on GAN
The GAN-based method addresses noise interference and channel sparsity in large-scale MIMO systems by optimizing input for channel estimation using a noise reduction neural network and attention mechanism, resulting in improved accuracy and efficiency.
Patent Information
- Application Number
- CN202310231727.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-03-10
- Publication Date
- 2025-07-15
- Estimated Expiration
- 2043-03-10
AI Technical Summary
In the existing large-scale MIMO systems, the impact of noise on the accuracy of channel estimation has not been fully considered, and the sparse structure has not been effectively utilized, resulting in high complexity of channel estimation, large power consumption, and long calculation time, making it difficult to meet the real-time processing requirements.
The GAN-based channel estimation method is used to remove noise through the denoising neural network, and the attention mechanism is used to focus on the local area of the channel to improve the accuracy and efficiency of channel estimation.
The impact of noise on channel estimation is reduced, and the efficiency and accuracy of channel learning is improved, especially in low signal-to-noise ratio conditions.
Smart Images

Figure CN116319190B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a channel estimation method, device, equipment and medium for a large-scale MIMO system based on GAN, belonging to the technical field of channel estimation in communication systems. Background Technique
[0002] Large-scale multiple-input multiple-output (MIMO) technology is one of the key technologies for fifth-generation (5G) and future cellular communication systems. By deploying a large number of antennas at the base station in a distributed or centralized manner, the large-scale MIMO system utilizes more spatial degrees of freedom than traditional MIMO systems, doubling the channel capacity without increasing the spectral resources and transmit power, and significantly reducing the interference between users by fully utilizing the spatial resources. In a large-scale MIMO system, accurate uplink and downlink channel state information (CSI) is crucial for signal detection, beamforming, resource allocation, signal preprocessing, etc. In the channel estimation of a multi-antenna system, to ensure the accuracy of channel estimation, the length of the pilot is usually greater than the number of transmit antennas, which will lead to significant pilot overhead when the number of antennas is large. The compressive sensing method can be used for channel estimation by utilizing the characteristics such as the sparse structure of the channel, but the complexity of this method increases with the increase in the number of antennas. With the use of large-scale MIMO, such methods have problems such as high complexity, high power consumption or large pilot overhead, and require a large amount of computing time and resources, and cannot meet the real-time processing requirements in actual deployment.
[0003] In recent years, deep learning (DL) technology has been widely applied in the field of wireless communication. In solving the problem of wireless channel estimation, deep learning provides an effective model for the channel estimation problem with its powerful function fitting ability. By utilizing this ability of deep learning, channel estimation accuracy can be guaranteed while having less pilot overhead and lower computational complexity.
[0004] Existing solutions for channel estimation using deep learning focus more on the improvement of the deep network itself, while ignoring the impact of noise on the channel estimation result. The noise mixed in the received signal will reduce the accuracy of channel estimation, and most of the existing solutions do not utilize the sparse structure of the channel itself.
[0005] It can be seen that in order to avoid the impact of noise mixed in the received signal on reducing the accuracy of channel estimation, there is an urgent need for a channel estimation method, device and storage medium for a large-scale MIMO system based on GAN. Summary of the Invention
[0006] The object of the present invention is to overcome the deficiencies in the prior art, and provide a channel estimation method, device, equipment and medium for a large-scale MIMO system based on GAN. By effectively denoising the received signal, the influence of noise on channel estimation in the channel estimation process is reduced, and the attention mechanism is adopted to effectively focus on each local area of the channel, improving the efficiency of channel learning and the accuracy of channel estimation.
[0007] To achieve the above object, the present invention is implemented by the following technical solutions:
[0008] In the first aspect, the present invention provides a channel estimation method for a large-scale MIMO system based on GAN, including the following steps:
[0009] Receive the pilot signal sent by the transmitting end of the MIMO system, and denoise the received signal based on the pre-designed and trained denoising neural network;
[0010] According to the signal obtained after denoising, use the stochastic gradient descent algorithm to optimize the random input of the pre-designed and trained generative adversarial network GAN to obtain the optimal GAN input;
[0011] Send the obtained optimal GAN input into the GAN generator, and obtain the optimal channel estimation value according to its output result;
[0012] Among them, the pre-designed and trained GAN includes an attention mechanism module, which is used to endow the GAN discriminator with the ability to focus on each local area of the channel.
[0013] Further, the denoising neural network adopts the structure of an autoencoder, including three encoder blocks and three decoder blocks, where
[0014] Each encoder block includes a convolutional layer, a normalization layer, and a LeakyReLU activation function layer connected in sequence, and each decoder block is composed of an upsampling layer, a convolutional layer, a normalization layer, and a LeakyReLU activation function layer connected in sequence;
[0015] Each convolutional layer is provided with 32 convolutional kernels with a size of 3×3;
[0016] A skip connection is provided between each encoder block and its corresponding decoder block.
[0017] Further, the GAN generator network includes a generator input layer, a generator hidden layer, and a generator output layer, where
[0018] The GAN generator network includes a generator input layer, a generator hidden layer, and a generator output layer connected in sequence, where
[0019] The generator input layer is the first layer of the GAN generator network. It includes a fully connected layer, which processes a low-dimensional vector using the LeakyReLU activation function and reshapes it into a three-dimensional vector.
[0020] The second, third, and fourth layers are the generator hidden layers respectively. Each generator hidden layer consists of an upsampling layer, a convolutional layer, a batch normalization layer, and a LeakyReLU activation function layer connected in sequence. Among them, the convolutional layer is a two-dimensional convolution with a kernel size of 5×5 and a stride of 1.
[0021] The fifth layer is the generator output layer, which uses a two-dimensional convolution with a kernel size of 2 and a stride of 2.
[0022] Furthermore, the GAN discriminator network includes a discriminator input layer, a discriminator hidden layer, and a discriminator output layer connected in sequence. Among them,
[0023] The discriminator input layer is the first layer of the GAN discriminator network. It includes a two-dimensional convolutional layer with a kernel size of 3×3 and a stride of 1, a LeakyReLU activation function layer, and an attention mechanism module connected in sequence.
[0024] The second, third, and fourth layers are the discriminator hidden layers respectively. Each discriminator hidden layer contains a two-dimensional convolutional layer with a kernel size of 3×3 and a stride of 1, a batch normalization layer, a LeakyReLU activation function layer, and an attention mechanism module connected in sequence.
[0025] The fifth layer is the discriminator output layer, which includes a fully connected layer.
[0026] Furthermore, the construction method of the attention mechanism module includes the following steps:
[0027] Perform global average pooling on the original feature matrix, compress each column of each feature matrix into a real number. This real number has a global receptive field, and the output dimension matches the number of columns of the input feature matrix.
[0028] Use a neural network containing three fully connected layers to generate weights for each feature column of the original feature matrix. Take the vector parameters of the output of the neural network as the importance weights of each feature column after feature selection, and weight the original features of the original feature matrix to complete the recalibration of the original feature matrix in the column dimension.
[0029] Furthermore, the training method of the denoising neural network includes the following steps:
[0030] Take the received signal as sample data and input it into the denoising neural network. Use the mean square error function between two independent noisy received signals as the cost function for training the network.
[0031] Update the network parameters of the deep neural network using the ADAM optimization algorithm and set the initial learning rate;
[0032] Use the sample data to perform offline training on the denoising neural network to obtain the trained denoising neural network.
[0033] Furthermore, the training method of the generative adversarial network includes the following steps:
[0034] Input the channel sample data into the generative adversarial network and learn the distribution of the real channel by means of the game between the generator and the discriminator.
[0035] Update the network parameters of the deep neural network using the ADAM optimization algorithm and set the initial learning rate;
[0036] Use the channel sample data to perform offline training on the generative adversarial network to obtain the trained generative adversarial network.
[0037] In a second aspect, the present invention provides a channel estimation device for a large-scale MIMO system based on GAN, including:
[0038] A signal denoising module: used to receive the pilot signal sent by the transmitting end of the MIMO system and denoise the received signal based on the pre-designed and trained denoising neural network;
[0039] A GAN input optimization module: used to optimize the random input of the pre-designed and trained generative adversarial network GAN according to the denoised signal using the stochastic gradient descent algorithm to obtain the optimal GAN input;
[0040] A channel estimation value generation module: used to send the obtained optimal GAN input into the GAN generator and obtain the optimal channel estimation value according to its output result;
[0041] Among them, the pre-designed and trained GAN includes an attention mechanism module for endowing the GAN discriminator with the ability to focus on each local area of the channel.
[0042] In a third aspect, the present invention provides a channel estimation device for a large-scale MIMO system based on GAN, including a processor and a storage medium;
[0043] The storage medium is used to store instructions;
[0044] The processor is used to operate according to the instructions to execute the steps of the method according to the first aspect.
[0045] In a fourth aspect, the present invention provides a computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, the steps of the method according to the first aspect are implemented.
[0046] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:
[0047] The method, device, equipment and medium for channel estimation of a large-scale MIMO system based on GAN provided by the present invention effectively denoise the received signal, reduce the influence of noise on channel estimation in the channel estimation process, and effectively utilize the spatial aggregation sparse structure of the angular domain channel by using the attention mechanism. The entire angular space is divided into multiple small angular regions, and these small regions are respectively focused on, filtering out the influence of irrelevant information on the channel learning of this region, improving the efficiency of channel learning and the accuracy of channel estimation. Description of the Drawings
[0048] Figure 1 It is a schematic flow chart of the method for channel estimation of a large-scale MIMO system based on GAN provided in Embodiment 1;
[0049] Figure 2 It is a schematic principle diagram of the method for channel estimation of a large-scale MIMO system based on GAN provided in Embodiment 1;
[0050] Figure 3 It is a structural diagram of the denoising network used in Embodiment 1;
[0051] Figure 4 It is a structural diagram of the attention mechanism used in Embodiment 1;
[0052] Figure 5 It is a comparison diagram of the channel normalized mean square error - signal-to-noise ratio curves of the three channel estimation methods: the GAN channel estimation method, the OMP channel estimation method, and the channel estimation method provided in Embodiment 1;
[0053] Figure 6 It is a schematic principle diagram of the device for channel estimation of a large-scale MIMO system based on GAN provided in Embodiment 2;
[0054] Figure 7 It is a specific process diagram implemented by the method for channel estimation of a large-scale MIMO system based on GAN provided in Embodiment 1. Detailed Embodiments
[0055] The technical solution of the present invention will be described in detail below through the drawings and specific embodiments. It should be understood that the specific features in the embodiments of the present application and the embodiments are detailed descriptions of the technical solution of the present application, rather than limitations on the technical solution of the present application. Without conflict, the technical features in the embodiments of the present application and the embodiments can be combined with each other.
[0056] In this article, the term "and / or" is merely a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can represent three situations: A exists alone, A and B exist simultaneously, and B exists alone. Additionally, in this article, the character " / " generally indicates that the associated objects before and after are in an "or" relationship.
[0057] Embodiment 1:
[0058] Figure 1 is a flowchart of a channel estimation method for a large-scale MIMO system based on GAN in Embodiment 1 of the present invention. The channel estimation method for a large-scale MIMO system based on GAN provided in this embodiment can be applied to a terminal and can be executed by a channel estimation device for a large-scale MIMO system based on GAN. This device can be implemented in a software and / or hardware manner and can be integrated into the terminal. For example: any smart phone, tablet computer or computer device with communication functions. Refer to Figure 1 , the method of this embodiment specifically includes the following steps:
[0059] Step A: Receive the pilot signal sent by the transmitting end of the MIMO system, and denoise the received signal based on a pre-designed and trained denoising neural network;
[0060] Step B: According to the signal obtained after denoising, use the stochastic gradient descent algorithm to optimize the random input of the pre-designed and trained generative adversarial network GAN to obtain the optimal GAN input;
[0061] Step C: Send the obtained optimal GAN input into the GAN generator, and obtain the optimal channel estimation value according to its output result;
[0062] Among them, the pre-designed and trained GAN includes an attention mechanism module for endowing the GAN discriminator with the ability to focus on each local area of the channel.
[0063] As an emerging generation framework, GAN has shown a powerful ability to generate simulated samples, that is, to capture the distribution of actual data, and has been used in the field of channel estimation. Denoising convolutional network (DnCNN), deep image prior (DIP), Noise2Noise, etc. are deep learning solutions commonly used in the field of image denoising and have achieved good denoising effects. Applying these denoising methods to the field of channel estimation can significantly improve the quality of the estimated channel. The attention mechanism has been widely applied in many fields of deep learning. Models based on the attention mechanism can not only record the positional relationship between information but also measure the importance of different information features according to the weight of the information. Reasonable application of these technologies in the field of channel estimation will achieve beneficial effects.
[0064] The denoising neural network adopts the structure of an autoencoder, including three encoder blocks and three decoder blocks. Among them,
[0065] Each encoder block contains a convolutional layer, a normalization layer, and a LeakyReLU activation function layer connected in sequence. Each decoder block consists of an upsampling layer, a convolutional layer, a normalization layer, and a LeakyReLU activation function layer connected in sequence;
[0066] Each convolutional layer is provided with 32 convolutional kernels of size 3×3;
[0067] A skip connection is set between each encoder block and its corresponding decoder block. Adding the skip connection enables the feature maps of the encoder block and the decoder block to be combined through a concatenation layer to retain the feature details at different resolutions, thereby retaining information at different scales.
[0068] The GAN generator network includes a generator input layer, a generator hidden layer, and a generator output layer connected in sequence. Among them,
[0069] The generator input layer is the first layer of the GAN generator network, which includes a fully connected layer, processes the low-dimensional vector using the LeakyReLU activation function, and reshapes it into a three-dimensional vector;
[0070] The second, third, and fourth layers are respectively the generator hidden layer modules. Each generator hidden layer module consists of an upsampling layer, a convolutional layer, a batch normalization layer, and a LeakyReLU activation function layer connected in sequence, where the convolutional layer is a two-dimensional convolution with a kernel size of 5×5 and a stride of 1;
[0071] The fifth layer is the generator output layer, which uses a two-dimensional convolution with a kernel size of 2 and a stride of 2. Upsampling increases the dimension so that the data output at the generator output end has the same size as the real channel matrix.
[0072] The GAN discriminator network includes a discriminator input layer, a discriminator hidden layer, and a discriminator output layer connected in sequence. Among them,
[0073] The discriminator input layer is the first layer of the GAN discriminator network, which includes a two-dimensional convolutional layer with a kernel size of 3×3 and a stride of 1, a LeakyReLU activation function layer, and an attention mechanism module connected in sequence;
[0074] The second, third, and fourth layers are respectively the discriminator hidden layers. Each discriminator hidden layer contains a two-dimensional convolutional layer with a kernel size of 3×3 and a stride of 1, a batch normalization layer, a LeakyReLU activation function layer, and an attention mechanism module connected in sequence;
[0075] The fifth layer is the discriminator output layer, which includes a fully connected layer. According to the mean value output by the output layer, the channel data output by the generator is classified as real or fake.
[0076] Since the quality of the channel distribution learned by the generator depends on the judgment of the discriminator, the ability of the discriminator to distinguish real channels from fake channels is crucial for whether the generator can capture real channel data with high quality. Therefore, the present invention adopts an attention mechanism module in the GAN to endow the discriminator with the ability to focus on each local area of the channel; the construction method of the attention mechanism module in this embodiment includes the following steps:
[0077] Compression operation: Perform global average pooling on the original feature matrix, compress each column of each feature matrix into a real number, this real number has a global receptive field, and the output dimension matches the number of columns of the input feature matrix;
[0078] Weight learning: Use a neural network containing three fully connected layers to generate weights for each feature column of the original feature matrix, and use the vector parameters output by the neural network as the importance weights of each feature column after feature selection, and weight the original features of the original feature matrix to complete the recalibration of the original feature matrix in the column dimension. The structural diagram of the attention mechanism is as Figure 4 shown.
[0079] The training method of the denoising neural network includes the following steps:
[0080] Take the received signal as sample data and input it into the denoising neural network, and use the mean square error function between two independent noisy received signals as the cost function for training the network. The expression is:
[0081]
[0082] In the formula, y n1 and y n2 are two received signals with independent noises, the function f is the mapping relationship between the input and output of the denoising neural network, T represents the number of training samples, and θ represents the weight parameters of the denoising neural network.
[0083] Use the ADAM optimization algorithm to update the network parameters of the deep neural network, set the initial learning rate, and in this embodiment, the initial learning rate is set to 0.01, epochs is set to 50, and bath size is set to 100;
[0084] Use the sample data to perform offline training on the denoising neural network, and deploy the trained network at the receiving end. In this embodiment, 1000 data samples are used as the training data set, and 50 data are used for testing.
[0085] The sample data of the denoising network is obtained by transmitting pilot signals twice within a time interval when the channel remains unchanged, and the received pilot signals are regarded as two independent noisy versions of the same pilot signal without noise. Enough signal samples for training the denoising network are generated by the above method..
[0086] The training method of the generative adversarial network includes the following steps:
[0087] Input the channel sample data into the generative adversarial network, and learn the distribution of the real channel by the way of the game between the generator and the discriminator. The cost function is as follows:
[0088] min G max D f(G, D) = E[h D (D(x; θ d ))] + E[h G (D(G(z; θ g )))]
[0089] Wherein,
[0090] h D (D(x; θ d )) = D(x; θ d )
[0091] h G (D(G(z; θ g ))) = -D(G(z; θ g ))
[0092] In the formula, G(z) represents the mapping from the input noise to the data space, D(x) represents the probability that x is the real channel data, and h D (D(x; θ d )) and h G (D(G(z; θ g )) represent the loss functions of the discriminator and the generator respectively. In this embodiment, the Wasserstein distance is used as the loss function, and using this loss function can reduce the occurrence of mode collapse.
[0093] Update the network parameters of the deep neural network using the ADAM optimization algorithm, and set the initial learning rate. In this embodiment, the initial learning rate is set to 0.001, epochs is set to 100, and bath size is set to 200;
[0094] Offline train the generative adversarial network using the channel sample data, and deploy the trained generator network at the receiving end. In this embodiment, the initial learning rate is set to 0.001, epochs is set to 100, and bath size is set to 200.
[0095] In this embodiment, the COST2100 model is adopted to generate the original channel matrix data as the channel sample data input when training the generative adversarial network. The scenario of 5.3 GHz is used. The base station is centered on a square area of 20 m × 20 m, and the user equipment moves randomly within the square. The base station adopts a uniform linear array, with the number of base station antennas being 256 and the number of users being 32.
[0096] The original channel matrix data is subjected to discrete Fourier transform, and the transformed sample data is used for offline training of the generative adversarial network.
[0097] By performing discrete Fourier transform on the channel matrix, an approximately sparse channel matrix in the angular domain is obtained. Define the channel H in the angular domain a as follows:
[0098] H a = F a H s F b
[0099] where F a and F b are discrete Fourier transform matrices, and H s is the original MIMO channel matrix.
[0100] First, the received signal at the receiving end is preprocessed for denoising to obtain the denoised signal y, which will be used in the process of optimizing the input of the generative network. Second, the following objective is solved using the stochastic gradient descent algorithm:
[0101] z * = argmin z f(y, XG(z))
[0102] for the optimal solution, where y is the denoised received signal obtained after passing through the denoising network, f represents the function of the two-norm, X is the measurement matrix, z * represents the optimal GAN input, and z represents the random noise sequence input to the GAN generator. The optimal random input noise sequence is obtained by solving using the stochastic gradient descent algorithm, and the output G(z) of the GAN generator corresponding to this optimal input sequence is the optimal channel estimate value.
[0103] To further verify the performance of the method in this embodiment, the performance of the channel estimation method in this embodiment is verified through simulation. Figure 5The contrast graph of the channel normalized mean square error - signal - to - noise ratio curves for three channel estimation methods: the GAN channel estimation method, the OMP channel estimation method, and the channel estimation method proposed in this paper. It can be seen that the channel estimation method proposed in this embodiment has better performance, especially excellent performance at low signal - to - noise ratios. This is because the denoising scheme proposed in this paper effectively denoises the received signal, reduces the influence of noise on channel estimation during the channel estimation process, and uses the attention mechanism to effectively utilize the spatial aggregation sparse structure of the angular - domain channel. The entire angular space is divided into multiple small angular regions, and these small regions are respectively focused on, filtering out the influence of irrelevant information on the channel learning of this region, and improving the efficiency and accuracy of channel learning.
[0104] In summary, the channel estimation method for large - scale MIMO systems based on GAN provided in this embodiment effectively denoises the received signal, reduces the influence of noise on channel estimation during the channel estimation process, and uses the attention mechanism to effectively utilize the spatial aggregation sparse structure of the angular - domain channel. The entire angular space is divided into multiple small angular regions, and these small regions are respectively focused on, filtering out the influence of irrelevant information on the channel learning of this region, and improving the efficiency of channel learning and the accuracy of channel estimation.
[0105] Embodiment 2:
[0106] As Figure 6 shown, this embodiment provides a channel estimation device for large - scale MIMO systems based on GAN, including:
[0107] A signal denoising module: used to receive the pilot signal sent by the transmitter of the MIMO system and denoise the received signal based on a pre - designed and trained denoising neural network;
[0108] A GAN input optimization module: used to optimize the random input of a pre - designed and trained generative adversarial network GAN according to the denoised signal using the stochastic gradient descent algorithm to obtain the optimal GAN input;
[0109] A channel estimation value generation module: used to send the obtained optimal GAN input into the GAN generator and obtain the optimal channel estimation value according to its output result;
[0110] Among them, the pre - designed and trained GAN includes an attention mechanism module, which is used to endow the GAN discriminator with the ability to focus on each local region of the channel.
[0111] The channel estimation device for large - scale MIMO systems based on GAN provided by the embodiments of the present invention can execute the channel estimation device method for large - scale MIMO systems based on GAN provided by any embodiment of the present invention, and has the corresponding functional modules and beneficial effects for executing the method.
[0112] Embodiment III:
[0113] The embodiment of the present invention further provides a channel estimation device for a large-scale MIMO system based on GAN, including a processor and a storage medium;
[0114] The storage medium is used for storing instructions;
[0115] The processor is configured to operate according to the instructions to execute the steps of the following method:
[0116] Receive the pilot signal sent by the transmitting end of the MIMO system, and denoise the received signal based on the pre-designed and trained denoising neural network;
[0117] According to the signal obtained after denoising, use the stochastic gradient descent algorithm to optimize the random input of the pre-designed and trained generative adversarial network GAN to obtain the optimal GAN input;
[0118] Send the obtained optimal GAN input into the GAN generator, and obtain the optimal channel estimation value according to its output result.
[0119] Embodiment IV:
[0120] The embodiment of the present invention further provides a computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, the steps of the following method are implemented:
[0121] Receive the pilot signal sent by the transmitting end of the MIMO system, and denoise the received signal based on the pre-designed and trained denoising neural network;
[0122] According to the signal obtained after denoising, use the stochastic gradient descent algorithm to optimize the random input of the pre-designed and trained generative adversarial network GAN to obtain the optimal GAN input;
[0123] Send the obtained optimal GAN input into the GAN generator, and obtain the optimal channel estimation value according to its output result.
[0124] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0125] This application is described with reference to the flowcharts and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the present application. It should be understood that each flow and / or block in the flowchart and / or block diagram, and the combination of flows and / or blocks in the flowchart and / or block diagram, can be realized by computer program instructions. These computer program instructions can be provided to the processors of general-purpose computers, special-purpose computers, embedded processors, or other programmable data processing devices to generate a machine, such that the instructions executed by the processors of the computer or other programmable data processing devices produce means for realizing the functions specified in the Figure 1 one or more flows and / or blocks Figure 1 means for realizing the functions specified in one or more blocks.
[0126] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, such that the instructions stored in the computer-readable memory produce a manufactured article including instruction means that realizes the functions specified in the Figure 1 one or more flows and / or blocks Figure 1 means for realizing the functions specified in one or more blocks.
[0127] These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process, so that the instructions executed on the computer or other programmable device provide steps for realizing the functions specified in the Figure 1 one or more flows and / or blocks Figure 1 means for realizing the functions specified in one or more blocks.
[0128] The above is only the preferred embodiment of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the technical principle of the present invention, several improvements and modifications can be made, and these improvements and modifications should also be regarded as the protection scope of the present invention.
Claims
1. A channel estimation method for a large-scale MIMO system based on GAN, characterized in that The method includes the following steps: Receiving a pilot signal sent by a transmitting end of an MIMO system, and denoising the received signal based on a pre-designed and pre-trained denoising neural network; According to the signal obtained after denoising, using the stochastic gradient descent algorithm to optimize the random input of a pre-designed and pre-trained generative adversarial network (GAN) to obtain an optimal GAN input; Feeding the obtained optimal GAN input into the GAN generator, and obtaining an optimal channel estimation value according to its output result; Wherein, the pre-designed and pre-trained GAN includes an attention mechanism module for endowing the GAN discriminator with the ability to focus on each local area of the channel; The denoising neural network adopts the structure of an autoencoder, and includes three encoder blocks and three decoder blocks, wherein, Each encoder block includes a convolutional layer, a normalization layer, and a LeakyReLU activation function layer connected in sequence, and each decoder block is composed of an upsampling layer, a convolutional layer, a normalization layer, and a LeakyReLU activation function layer connected in sequence; Each convolutional layer is provided with 32 convolutional kernels with a size of 3×3; A skip connection is provided between each encoder block and its corresponding decoder block.
2. The channel estimation method for a large-scale MIMO system based on GAN according to claim 1, wherein The generator network of the GAN includes a generator input layer, a generator hidden layer, and a generator output layer connected in sequence, wherein, The generator input layer is the first layer of the generator network of the GAN, and includes a fully connected layer, which processes a low-dimensional vector using a LeakyReLU activation function and reshapes it into a three-dimensional vector; The second, third, and fourth layers are respectively the generator hidden layers, and each generator hidden layer is composed of an upsampling layer, a convolutional layer, a batch normalization layer, and a LeakyReLU activation function layer connected in sequence, wherein the convolutional layer is a two-dimensional convolution with a convolutional kernel size of 5×5 and a stride of 1; The fifth layer is the generator output layer, which uses a two-dimensional convolution with a convolutional kernel size of 2 and a stride of 2.
3. The channel estimation method for a large-scale MIMO system based on GAN according to claim 1, characterized in that, The discriminator network of the GAN includes a discriminator input layer, a discriminator hidden layer, and a discriminator output layer connected in sequence, wherein, The discriminator input layer is the first layer of the discriminator network of the GAN, and includes a two-dimensional convolutional layer with a convolutional kernel size of 3×3 and a stride of 1, a LeakyReLU activation function layer, and an attention mechanism module connected in sequence; The second, third, and fourth layers are respectively the discriminator hidden layers, and each discriminator hidden layer includes a two-dimensional convolutional layer with a convolutional kernel size of 3×3 and a stride of 1, a batch normalization layer, a LeakyReLU activation function layer, and an attention mechanism module connected in sequence; The fifth layer is the discriminator output layer, which includes a fully connected layer.
4. The channel estimation method for a large-scale MIMO system based on GAN according to claim 1, characterized in that, The construction method of the attention mechanism module includes the following steps: Performing global average pooling on the original feature matrix, compressing each column of each feature matrix into a real number, this real number has a global receptive field, and the output dimension matches the number of columns of the input feature matrix; A neural network with three fully-connected layers is used to generate weights for each feature column of the original feature matrix. The vector parameters of the output of the neural network are used as the importance weights of each feature column after feature selection, and the original features of the original feature matrix are weighted to complete the recalibration of the original feature matrix in the column dimension.
5. The channel estimation method for a large-scale MIMO system based on GAN according to claim 1, characterized in that, The training method of the denoising neural network includes the following steps: The received signal is used as sample data and input into the denoising neural network, and the mean square error function between two independent noisy received signals is used as the cost function for training the network. The network parameters of the deep neural network are updated using the ADAM optimization algorithm, and the initial learning rate is set. The denoising neural network is offline-trained using the sample data to obtain the trained denoising neural network.
6. The channel estimation method for a large-scale MIMO system based on GAN according to claim 1, characterized in that, The training method of the generative adversarial network includes the following steps: The channel sample data is input into the generative adversarial network, and the distribution of the real channel is learned by the way of the game between the generator and the discriminator. The network parameters of the deep neural network are updated using the ADAM optimization algorithm, and the initial learning rate is set. The generative adversarial network is offline-trained using the channel sample data to obtain the trained generative adversarial network.
7. A channel estimation device for a large-scale MIMO system based on GAN, characterized in that, It includes: Signal denoising module: Receives the pilot signal sent by the transmitting end of the MIMO system, and denoises the received signal based on the pre-designed and trained denoising neural network. GAN input optimization module: Used to optimize the random input of the pre-designed and trained generative adversarial network GAN using the stochastic gradient descent algorithm according to the signal obtained after denoising to obtain the optimal GAN input. Channel estimation value generation module: Used to send the obtained optimal GAN input into the GAN generator and obtain the optimal channel estimation value according to its output result. Among them, the pre-designed and trained GAN includes an attention mechanism module, which is used to endow the GAN discriminator with the ability to focus on each local area of the channel. The denoising neural network adopts the structure of an autoencoder, including three encoder blocks and three decoder blocks, where Each encoder block contains a convolutional layer, a normalization layer, and a LeakyReLU activation function layer connected in sequence, and each decoder block is composed of an upsampling layer, a convolutional layer, a normalization layer, and a LeakyReLU activation function layer connected in sequence; Each convolutional layer is provided with 32 convolutional kernels of size 3×3; A skip connection is provided between each encoder block and its corresponding decoder block.
8. A channel estimation device for a large-scale MIMO system based on GAN, characterized in that It includes a processor and a storage medium; The storage medium is used to store instructions; The processor is used to operate according to the instructions to execute the steps of the method according to any one of claims 1 to 6.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it realizes the steps of the method according to any one of claims 1 to 6.
Citation Information
Patent Citations
Information generation method and related device
CN114666188A
Image deblurring algorithm based on knowledge distillation and deep neural network
CN114677304A